Gradient methods for constrained problemsyc5/ele522_optimization/lectures/grad_descent... · Gradient methods for constrained problems Yuxin Chen Princeton University, Fall 2019.

ELE 522: Large-Scale Optimization for Data Science

Gradient methods for constrained problems

Yuxin Chen

Princeton University, Fall 2019

Outline

• Frank-Wolfe algorithm

• Projected gradient methods

Gradient methods (constrained case) 3-2

Constrained convex problems

minimizex f(x)subject to x ∈ C

• f(·): convex function• C ⊆ Rn: closed convex set


Feasible direction methods

Generate a feasible sequence xt ⊆ C with iterations

xt+1 = xt + ηtdt

where dt is a feasible direction (s.t. xt + ηtdt ∈ C)

• Question: can we guarantee feasibility while enforcing costimprovement?


Frank-Wolfe algorithm

Frank-Wolfe algorithm was developed by Philip Wolfe and MargueriteFrank when they worked at / visited Princeton

Frank-Wolfe / conditional gradient algorithm

Algorithm 3.1 Frank-wolfe (a.k.a. conditional gradient) algorithm1: for t = 0, 1, · · · do2: yt := arg minx∈C 〈∇f(xt),x〉 (direction finding)3: xt+1 = (1− ηt)xt + ηty

t (line search and update)

yt = arg minx∈C〈∇f(xt),x− xt〉

Frank-Wolfe/ conditional gradient algorithm

Algorithm 2.3 Frank-wolfe (a.k.a. conditional gradient) algorithm1: for t = 0, 1, · · · do2: yt := arg minxœX ÈÒf(xt),x ≠ xtÍ (direction finding)3: xt+1 = (1 ≠ ÷t)xt + ÷ty


• main step: linearization of objective function (same asf(xt) + ÈÒf(xt),x ≠ xtÍ)

=∆ linear optimization over convex set• appealing when linear optimization is cheap• stepsize ÷t determined by line search, or ÷t = 2

t+1

Gradient methods 2-44

Monotonicity

We start with a monotonicity result:

Lemma 2.5

Let f be convex and L-smooth. If ÷t © ÷ = 1/L, then

Îxt+1 ≠ xúÎ2 Æ Îxt ≠ xúÎ2

where xú is any minimizer with optimal f(xú)


Proof of Lemma 2.5

It follows that

Îxt+1 ≠ xúÎ22 =

..xt ≠ xú ≠ ÷(Òf(xt) ≠ Òf(xú)¸ ˚˙ ˝=0

)..2

2

=..xt ≠ xú..2

2 ≠ 2÷Èxt ≠ xú, Òf(xt) ≠ Òf(xú)Í¸ ˚˙ ˝Ø 2÷

L ÎÒf(xt)≠Òf(xú)Î22 (smoothness)

+ ÷2..Òf(xt) ≠ Òf(xú)..2

2

Æ..xt ≠ xú..2

2 ≠ ÷2..Òf(xt) ≠ Òf(xú)..2

2

Æ..xt ≠ xú..2

2


Proof of Lemma 2.5

It follows that


..xt ≠ xú ≠ ÷(Òf(xt) ≠ Òf(xú)¸ ˚˙ ˝=0

)..2

2

=..xt ≠ xú..2



+ ÷2..Òf(xt) ≠ Òf(xú)..2

2

Æ..xt ≠ xú..2

2 ≠ ÷2..Òf(xt) ≠ Òf(xú)..2

2

Æ..xt ≠ xú..2

2


The claim would follow immediately if

(x x)>(z z) kz zk2 (together with Cauchy-Schwarz)

(= (x z x + z)>(z z) 0

(=

8>>><>>>:

h(z) h (z) + hx z| z 2@h(z)

, z zi

h(z) h(z) + hx z| z 2@h(z)

, z zi

1

2k 1k2 + c1 1

2k 2k2 + c2 1 2 prox(1) prox(2)

relative errorf(t) + g(t) minf() + g()minf() + g()

iteration t

=

11 12

>12 22

S =

S11 s12

s>12 s22

W =

W11 w12

w>12 w22

0 2 W11 s12 + @k12k1

C PC(1) PC(2)

4

• main step: linearization of the objective function (equivalent tof(xt) + 〈∇f(xt),x− xt〉)

=⇒ linear optimization over a convex set• appealing when linear optimization is cheap

• stepsize ηt determined by line search, or ηt = 2t+ 2︸︷︷︸

bias towards xt for large t


Frank-Wolfe / conditional gradient algorithm

Algorithm 3.2 Frank-wolfe (a.k.a. conditional gradient) algorithm1: for t = 0, 1, · · · do2: yt := arg minx∈C 〈∇f(xt),x〉 (direction finding)3: xt+1 = (1− ηt)xt + ηty


yt = arg minx∈C〈∇f(xt),x− xt〉

Frank-Wolfe/ conditional gradient algorithm

Algorithm 2.3 Frank-wolfe (a.k.a. conditional gradient) algorithm1: for t = 0, 1, · · · do2: yt := arg minxœX ÈÒf(xt),x ≠ xtÍ (direction finding)3: xt+1 = (1 ≠ ÷t)xt + ÷ty


• main step: linearization of objective function (same asf(xt) + ÈÒf(xt),x ≠ xtÍ)

=∆ linear optimization over convex set• appealing when linear optimization is cheap• stepsize ÷t determined by line search, or ÷t = 2

t+1


Monotonicity

We start with a monotonicity result:

Lemma 2.5

Let f be convex and L-smooth. If ÷t © ÷ = 1/L, then

Îxt+1 ≠ xúÎ2 Æ Îxt ≠ xúÎ2

where xú is any minimizer with optimal f(xú)


Proof of Lemma 2.5

It follows that


..xt ≠ xú ≠ ÷(Òf(xt) ≠ Òf(xú)¸ ˚˙ ˝=0

)..2

2

=..xt ≠ xú..2



+ ÷2..Òf(xt) ≠ Òf(xú)..2

2

Æ..xt ≠ xú..2

2 ≠ ÷2..Òf(xt) ≠ Òf(xú)..2

2

Æ..xt ≠ xú..2

2


Proof of Lemma 2.5

It follows that


..xt ≠ xú ≠ ÷(Òf(xt) ≠ Òf(xú)¸ ˚˙ ˝=0

)..2

2

=..xt ≠ xú..2



+ ÷2..Òf(xt) ≠ Òf(xú)..2

2

Æ..xt ≠ xú..2

2 ≠ ÷2..Òf(xt) ≠ Òf(xú)..2

2

Æ..xt ≠ xú..2

2




(= (x z x + z)>(z z) 0

(=

8>>><>>>:

h(z) h (z) + hx z| z 2@h(z)

, z zi

h(z) h(z) + hx z| z 2@h(z)

, z zi

1

2k 1k2 + c1 1

2k 2k2 + c2 1 2 prox(1) prox(2)


iteration t

=

11 12

>12 22

S =

S11 s12

s>12 s22

W =

W11 w12

w>12 w22

0 2 W11 s12 + @k12k1

C PC(1) PC(2)

4

• main step: linearization of the objective function (equivalent tof(xt) + 〈∇f(xt),x− xt〉)

=⇒ linear optimization over a convex set• appealing when linear optimization is cheap

• stepsize ηt determined by line search, or ηt = 2t+ 2︸︷︷︸

bias towards xt for large t


Frank-Wolfe can also be applied to nonconvexproblems

Example (Luss & Teboulle ’13)

minimizex − x>Qx subject to ‖x‖2 ≤ 1 (3.1)

for some Q 0


Frank-Wolfe can also be applied to nonconvexproblems

We now apply Frank-Wolfe to solve (3.1). Clearly,

yt = arg minx:‖x‖2≤1

〈∇f(xt),x〉 = − ∇f(xt)‖∇f(xt)‖2

= Qxt

‖Qxt‖2

=⇒ xt+1 = (1− ηt)xt + ηtQxt/‖Qxt‖2Set ηt = arg min0≤η≤1 f

((1− η)xt + η Qxt

‖Qxt‖2

)= 1 (check). This

givesxt+1 = Qxt/‖Qxt‖2

which is essentially power method for finding leading eigenvector of Q


Convergence for convex and smooth problems

Theorem 3.1 (Frank-Wolfe for convex and smooth problems,Jaggi ’13)

Let f be convex and L-smooth. With ηt = 2t+2 , one has

f(xt)− f(x∗) ≤ 2Ld2C

t+ 2

where dC = supx,y∈C ‖x− y‖2

• for compact constraint sets, Frank-Wolfe attains ε-accuracywithin O(1

ε ) iterations


Proof of Theorem 3.1By smoothness,

f(xt+1)− f(xt) ≤ ∇f(xt)>(xt+1 − xt︸︷︷︸=ηt(yt−xt)

) + L

2 ‖xt+1 − xt‖22︸︷︷︸

=η2t ‖yt−xt‖2

2≤η2t d

2C

≤ ηt∇f(xt)>(yt − xt) + L

2 η2t d

2C

≤ ηt∇f(xt)>(x∗ − xt) + L

2 η2t d

2C (since yt is minimizer)

≤ ηt(f(x∗)− f(xt)

)+ L

2 η2t d

2C (by convexity)

Letting ∆t := f(xt)− f(x∗) we get

∆t+1 ≤ (1− ηt)∆t + Ld2C

2 η2t

We then complete the proof by induction (which we omit here)Gradient methods (constrained case) 3-10

Strongly convex problems?

Can we hope to improve convergence guarantees of Frank-Wolfe inthe presence of strong convexity?

• in general, NO• maybe improvable under additional conditions


A negative result

Example:

minimizex∈Rn12x>Qx + b>x (3.2)

subject to x = [a1, · · · ,ak]v, v ≥ 0, 1>v = 1︸︷︷︸x∈ convex-hulla1,··· ,ak

(=: Ω)

• suppose interior(Ω) 6= ∅• suppose the optimal point x∗ lies on the boundary of Ω and is

not an extreme point


A negative result

Theorem 3.2 (Canon & Cullum, ’68)Let xt be Frank-Wolfe iterates with exact line search for solving(3.2). Then ∃ an initial point x0 s.t. for every ε > 0,

f(xt)− f(x∗) ≥ 1t1+ε for infinitely many t

• example: choose x0 ∈ interior(Ω) obeying f(x0) < mini f(ai)• in general, cannot improve O(1/t) convergence guarantees


Positive results?

To achieve faster convergence, one needs additional assumptions

• example: strongly convex feasible set C• active research topics


An example of positive results

A set C is said to be µ-strongly convex if ∀λ ∈ [0, 1] and ∀x, z ∈ C:

B(λx + (1− λ)z, µ2λ(1− λ)‖x− z‖22

)∈ C,

where B(a, r) := y | ‖y − a‖2 ≤ r• example: `2 ball

Theorem 3.3 (Levitin & Polyak ’66)Suppose f is convex and L-smooth, and C is µ-strongly convex.Suppose ‖∇f(x)‖2 ≥ c > 0 for all x ∈ C. Then under mildconditions, Frank-Wolfe with exact line search converges linearly


Projected gradient methods

Projected gradient descent

x0 x1 x2 x3


x0 x1 x2 x3


x0 x1 x2 x3


x0 x1 x2 x3




(= (x z x + z)>(z z) 0

(=

8>>><>>>:

h(z) h (z) + hx z| z 2@h(z)

, z zi

h(z) h(z) + hx z| z 2@h(z)

, z zi

1

2k 1k2 + c1 1

2k 2k2 + c2 1 2 prox(1) prox(2)


iteration t

=

11 12

>12 22

S =

S11 s12

s>12 s22

W =

W11 w12

w>12 w22

0 2 W11 s12 + @k12k1

C PC(1) PC(2)

4

works well if projectiononto C can be

computed efficiently

for t = 0, 1, · · · :xt+1 = PC(xt − ηt∇f(xt))

where PC(x) := arg minz∈C ‖x− z‖22 is Euclidean projection︸︷︷︸quadratic minimization

onto C


Descent direction

Descent direction

xC = PC(x)

Nonexpansiveness of projection operator

PC(x) PC(z) x z

Gradient methods (constrained) 3-13

Descent direction

Fact 3.2 (Projection theorem)

Let C be closed convex set. Then xC is projection of x onto C i

(x ≠ xC)€(z ≠ xC) Æ 0, ’z œ C




PC(x) PC(z) x z



(x x)(z z) z z2 (together with Cauchy-Schwarz)

= (x z x + z)(z z) 0

=

h(z) h (z) + x z h(z)

, z z

h(z) h(z) + x z h(z)

, z z

1

2 12 + c1 1

2 22 + c2 1 2 prox(1) prox(2)


iteration t

=

11 12

12 22

S =

S11 s12

s12 s22

W =

W11 w12

w12 w22

0 W11 s12 + 121

C PC(1) PC(2)

4


PC(x) PC(z) x z



PC(x) PC(z) x z



PC(x) PC(z) x z


Fact 3.3 (Nonexpansivness of projection)

For any x and z, one has ÎPC(x) ≠ PC(z)Î2 Æ Îx ≠ zÎ2



PC(x) PC(z) x z



Let C be closed convex set. Then xC is projection of x onto C i(x ≠ xC)€(z ≠ xC) Æ 0, ’z œ C



PC(x) PC(z) x z


Descent direction



(x ≠ xC)€(z ≠ xC) Æ 0, ’z œ C



PC(x) PC(z) x z



Let C be closed & convex. Then xC is the projection of x onto C iff(x− xC)>(z − xC) ≤ 0, ∀z ∈ C


Descent directionGradient descent (GD)

One of most important examples of (2.3): gradient descent

xt+1 = xt ≠ ÷tÒf(xt) (2.3)

• traced to Augustin Louis Cauchy ’1847 ...

• descent direction: dt = ≠Òf(xt)• a.k.a. steepest descent, since from (2.1) and Cauchy-Schwarz,

arg mind:ÎdÎ2Æ1

f Õ(x;d) = ≠ Òf(x)ÎÒf(x)Î2¸ ˚˙ ˝

direction with greatest rate of cost improvement

Gradient methods (unconstrained) 2-5

Gradient descent (GD)

One of most important examples of (2.3): gradient descent

xt+1 = xt ≠ ÷tÒf(xt) (2.3)

• traced to Augustin Louis Cauchy ’1847 ...

• descent direction: dt = ≠Òf(xt)• a.k.a. steepest descent, since from (2.1) and Cauchy-Schwarz,

arg mind:ÎdÎ2Æ1

f Õ(x;d) = ≠ Òf(x)ÎÒf(x)Î2¸ ˚˙ ˝

direction with greatest rate of cost improvement

Gradient methods (unconstrained) 2-5

Descent direction

Let yt = xt ≠ ÷tÒf(xt) be gradient update before projection. ThenFact 3.2 implies

Òf(xt)€(xt+1 ≠ xt) = ÷≠1t (xt ≠ yt)€(xt ≠ xt+1) Æ 0

xt+1 = PC(yt)


Descent direction

Let yt = xt ≠ ÷tÒf(xt) be gradient update before projection. ThenFact 3.2 implies

Òf(xt)€(xt+1 ≠ xt) = ÷≠1t (xt ≠ yt)€(xt ≠ xt+1) Æ 0


Descent direction



(x ≠ xC)€(z ≠ xC) Æ 0, ’z œ C


From the above figure, we know

−∇f(xt)>(xt+1 − xt) ≥ 0

xt+1 − xt is positively correlated with the steepest descent direction


Strongly convex and smooth problems


• f(·): µ-strongly convex and L-smooth• C ⊆ Rn: closed and convex


Convergence for strongly convex and smoothproblems

Proof of Lemma 2.5

It follows that


..xt ≠ xú ≠ ÷(Òf(xt) ≠ Òf(xú)¸ ˚˙ ˝=0

)..2

2

=..xt ≠ xú..2



+ ÷2..Òf(xt) ≠ Òf(xú)..2

2

Æ..xt ≠ xú..2

2 ≠ ÷2..Òf(xt) ≠ Òf(xú)..2

2

Æ..xt ≠ xú..2

2


Let’s start with the simple case when x∗ lies in the interior of C (sothat ∇f(x∗) = 0)



Theorem 3.5

Suppose x∗ ∈ int(C), and let f be µ-strongly convex and L-smooth.If ηt = 2

µ+L , then

‖xt − x∗‖2 ≤(κ− 1κ+ 1

)t‖x0 − x∗‖2

where κ = L/µ is condition number

• the same convergence rate as for the unconstrained case


Aside: nonexpansiveness of projection operator


PC(x) PC(z) x z




(= (x z x + z)>(z z) 0

(=

8>>><>>>:

h(z) h (z) + hx z| z 2@h(z)

, z zi

h(z) h(z) + hx z| z 2@h(z)

, z zi

1

2k 1k2 + c1 1

2k 2k2 + c2 1 2 prox(1) prox(2)


iteration t

=

11 12

>12 22

S =

S11 s12

s>12 s22

W =

W11 w12

w>12 w22

0 2 W11 s12 + @k12k1

C PC(1) PC(2)

4


PC(x) PC(z) x z



PC(x) PC(z) x z



PC(x) PC(z) x z


Fact 3.6 (Nonexpansivness of projection)

For any x and z, one has ‖PC(x)− PC(z)‖2 ≤ ‖x− z‖2


Proof of Theorem 3.5

We have shown for the unconstrained case that

‖xt − ηt∇f(xt)− x∗‖2 ≤κ− 1κ+ 1‖x

t − x∗‖2

From the nonexpansiveness of PC , we know

‖xt+1 − x∗‖2 = ‖PC(xt − ηt∇f(xt))− PC(x∗)‖2≤ ‖xt − ηt∇f(xt)− x∗‖2≤ κ− 1κ+ 1‖x

t − x∗‖2

Apply it recursively to conclude the proof



Proof of Lemma 2.5

It follows that


..xt ≠ xú ≠ ÷(Òf(xt) ≠ Òf(xú)¸ ˚˙ ˝=0

)..2

2

=..xt ≠ xú..2



+ ÷2..Òf(xt) ≠ Òf(xú)..2

2

Æ..xt ≠ xú..2

2 ≠ ÷2..Òf(xt) ≠ Òf(xú)..2

2

Æ..xt ≠ xú..2

2


What happens if we don’t know whether x∗ ∈ int(C)?• main issue: ∇f(x∗) may not be 0 (so prior analysis might fail)



Theorem 3.7 (projected GD for strongly convex and smoothproblems)

Let f be µ-strongly convex and L-smooth. If ηt ≡ η = 1L , then

‖xt − x∗‖22 ≤(

1− µ

L

)t‖x0 − x∗‖22

• slightly weaker convergence guarantees than Theorem 3.5


Proof of Theorem 3.7Let x+ := PC(x− 1

L∇f(x)) and gC(x) := 1η (x− x+) = L(x− x+)

︸︷︷︸negative descent direction

• gC(x) generalizes ∇f(x) and obeys gC(x∗) = 0

Main pillar:

〈gC(x),x− x∗〉 ≥ µ

2 ‖x− x∗‖22 + 12L‖gC(x)‖22 (3.3)

• this generalizes the regularity condition for GD

With (3.3) in place, repeating GD analysis under the regularitycondition gives

‖xt+1 − x∗‖22 ≤(

1− µ

L

)‖xt − x∗‖22

which immediately establishes Theorem 3.7Gradient methods (constrained case) 3-27

Proof of Theorem 3.7 (cont.)It remains to justify (3.3). To this end, it is seen that

0 ≤ f(x+)− f(x∗) = f(x+)− f(x) + f(x)− f(x∗)

≤ ∇f(x)>(x+ − x) + L

2 ‖x+ − x‖2

2︸︷︷︸

smoothness

+∇f(x)>(x− x∗)− µ

2 ‖x− x∗‖22

︸︷︷︸strong convexity

= ∇f(x)>(x+ − x∗) + 12L‖gC(x)‖2

2 −µ

2 ‖x− x∗‖22,

which would establish (3.3) if

∇f(x)>(x+−x∗) ≤ gC(x)>(x+ − x∗)︸︷︷︸=gC(x)>(x−x∗)− 1

L‖gC(x)‖22

(projection only makes it better)

(3.4)This inequality is equivalent to

(x+ −

(x− L−1∇f(x)

))> (x+ − x∗) ≤ 0 (3.5)

This fact (3.5) follows directly from Fact 3.4Gradient methods (constrained case) 3-28

RemarkGeneralized Pythagorean Theorem

[PICTURE]

Fact 5.2

If xC = PC,Ï(x), thenDÏ(z,x) Ø DÏ(z,xC) + DÏ(xC ,x) ’z œ C

• in squared Euclidean case, it means angle \zxCx is obtuse• if C is ane plane, then

DÏ(z,x) = DÏ(z,xC) + DÏ(xC ,x) ’z œ C

Mirror descent 5-26

Three-point lemma

[PICTURE]

Fact 5.1

For every three points x,y,z,

DÏ(x,z) = DÏ(x,y) + DÏ(y,z) ≠ ÈÒÏ(z) ≠ ÒÏ(y),x ≠ yÍ

• for Euclidean case with Ï(x) = ÎxÎ22, this is law of cosine

Îx ≠ zÎ22 = Îx ≠ yÎ2

2 + Îy ≠ zÎ22 ≠ 2 Èz ≠ y,x ≠ yÍ¸ ˚˙ ˝

Îz≠yÎ2Îx≠yÎ2 cos\zyx

Mirror descent 5-21

rn

p. minp

4µr log n

Since kM?k1 µrn min, we have

minp4µr log n

nkM?k1µr

p4µr log n

Therefore, we have

rn

p. nkM?k1

µrp4µr log n

() .r

np

6µ3r3 log nkM?k1

x 1Lrf(x)

1

rn

p. minp

4µr log n


minp4µr log n

nkM?k1µr

p4µr log n

Therefore, we have

rn

p. nkM?k1

µrp4µr log n

() .r

np

6µ3r3 log nkM?k1

x 1Lrf(x) x+

1

rn

p. minp

4µr log n


minp4µr log n

nkM?k1µr

p4µr log n

Therefore, we have

rn

p. nkM?k1

µrp

4µr log n

() .r

np

6µ3r3 log nkM?k1

x 1Lrf(x) x+ x

1

rn

p. minp

4µr log n


minp4µr log n

nkM?k1µr

p4µr log n

Therefore, we have

rn

p. nkM?k1

µrp

4µr log n

() .r

np

6µ3r3 log nkM?k1

x 1Lrf(x) x+ x y

1

One can easily generalize (3.4) to (via the same proof)

∇f(x)>(x+ − y) ≤ gC(x)>(x+ − y), ∀y ∈ C (3.6)

This proves useful for subsequent analysisGradient methods (constrained case) 3-29

Convex and smooth problems


• f(·): convex and L-smooth• C ⊆ Rn: closed and convex


Convergence for convex and smooth problems

Theorem 3.8 (projected GD for convex and smooth problems)

Let f be convex and L-smooth. If ηt ≡ η = 1L , then

f(xt)− f(x∗) ≤ 3L‖x0 − x∗‖22 + f(x0)− f(x∗)t+ 1

• similar convergence rate as for the unconstrained case


Proof of Theorem 3.8

We first recall our main steps when handling the unconstrained case

Step 1: show cost improvement

f(xt+1) ≤ f(xt)− 12L‖∇f(xt)‖2

2

Step 2: connect ‖∇f(xt)‖2 with f(xt)

‖∇f(xt)‖2 ≥f(xt)− f(x∗)‖xt − x∗‖2

≥ f(xt)− f(x∗)‖x0 − x∗‖2

Step 3: let ∆t := f(xt)− f(x∗) to get

∆t+1 −∆t ≤ −∆2t

2L‖x0 − x∗‖22

and complete the proof by inductionGradient methods (constrained case) 3-32

Proof of Theorem 3.8 (cont.)We then modify these steps for the constrained case. As before, setgC(xt) = L(xt − xt+1), which generalizes ∇f(xt) in constrained case

Step 1: show cost improvement

f(xt+1) ≤ f(xt)− 12L‖gC(x

t)‖22

Step 2: connect ‖gC(xt)‖2 with f(xt)

‖gC(xt)‖2 ≥f(xt+1)− f(x∗)‖xt − x∗‖2

≥ f(xt+1)− f(x∗)‖x0 − x∗‖2

Step 3: let ∆t := f(xt)− f(x∗) to get

∆t+1 −∆t ≤ −∆2t+1

2L‖x0 − x∗‖22

and complete the proof by inductionGradient methods (constrained case) 3-33

Proof of Theorem 3.8 (cont.)

Main pillar: generalize smoothness condition as follows

Lemma 3.9

Suppose f is convex and L-smooth. For any x,y ∈ C, letx+ = PC(x− 1

L∇f(x)) and gC(x) = L(x− x+). Then

f(y) ≥ f(x+) + gC(x)>(y − x) + 12L‖gC(x)‖22


Proof of Theorem 3.8 (cont.)Step 1: set x = y = xt in Lemma 3.9 to reach

f(xt) ≥ f(xt+1) + 12L‖gC(x

t)‖22

as desired

Step 2: set x = xt and y = x∗ in Lemma 3.9 to get

0 ≥ f(x∗)− f(xt+1) ≥ gC(xt)>(x∗ − xt) + 12L‖gC(x

t)‖22≥ gC(xt)>(x∗ − xt)

which together with Cauchy-Schwarz yields

‖gC(xt)‖2 ≥f(xt+1)− f(x∗)‖xt − x∗‖2

(3.7)


Proof of Theorem 3.8 (cont.)

It also follows from our analysis for the strongly convex case that (bytaking µ = 0 in Theorem 3.7)

‖xt − x∗‖2 ≤ ‖x0 − x∗‖2

which combined with (3.7) reveals

‖gC(xt)‖2 ≥f(xt+1)− f(x∗)‖x0 − x∗‖2

Step 3: letting ∆t = f(xt)− f(x∗), the previous bounds togethergive

∆t+1 −∆t ≤ −∆2t+1

2L‖x0 − x∗‖22Use induction to finish the proof (which we omit here)


Proof of Lemma 3.9

f(y)− f(x+) = f(y)− f(x)−(f(x+)− f(x)

)

≥ ∇f(x)>(y − x)︸︷︷︸convexity

−(∇f(x)>(x+ − x) + L

2 ‖x+ − x‖2

2

)

︸︷︷︸smoothness

= ∇f(x)>(y − x+)− L

2 ‖x+ − x‖2

2

≥ gC(x)>(y − x+)− L

2 ‖x+ − x‖2

2 (by (3.6))

= gC(x)>(y − x) + gC(x)>(x− x+)︸︷︷︸= 1

LgC(x)

− L

2 ‖ x+ − x︸︷︷︸=− 1

LgC(x)

‖22

= gC(x)>(y − x) + 12L‖gC(x)‖2

2


Summary

• Frank-Wolfe: projection-free

stepsize convergence iterationrule rate complexity

convex & smoothηt 1

t O( 1t

)O( 1ε

)problems

• projected gradient descent

stepsize convergence iterationrule rate complexity

convex & smoothηt = 1

L O( 1t

)O( 1ε

)problems

strongly convex &ηt = 1

L O((

1− 1κ

)t)O(κ log 1

ε

)smooth problems


Reference

[1] ”Nonlinear programming (3rd edition),” D. Bertsekas, 2016.[2] ”Convex optimization: algorithms and complexity,” S. Bubeck,

Foundations and trends in machine learning, 2015.[3] ”First-order methods in optimization,” A. Beck, Vol. 25, SIAM, 2017.[4] ”Convex optimization and algorithms,” D. Bertsekas, 2015.[5] ”Conditional gradient algorithmsfor rank-one matrix approximations with

a sparsity constraint,” R. Luss, M. Teboulle, SIAM Review, 2013.[6] ”Revisiting Frank-Wolfe: projection-free sparse convex optimization,”

M. Jaggi, ICML, 2013.[7] ”A tight upper bound on the rate of convergence of Frank-Wolfe

algorithm,” M. Canon and C. Cullum, SIAM Journal on Control, 1968.


Reference

[8] ”Constrained minimization methods,” E. Levitin and B. Polyak, USSRComputational mathematics and mathematical physics, 1966.


Gradient methods for constrained problemsyc5/ele522_optimization/lectures/grad_descent... · Gradient methods for constrained problems Yuxin Chen Princeton University, Fall 2019.

Documents